POLONICI MATHEMATICI LXXV.2 (2000)
On Cohen’s proof of the Factorization Theorem
by Jan Kisy´ nski (Lublin)
Dedicated to the memory of Mieczys law Altman
Abstract. Various proofs of the Factorization Theorem for representations of Banach algebras are compared with its original proof due to P. Cohen.
Various proofs of the Cohen–Hewitt Factorization Theorem for represen- tations of Banach algebras are presented and discussed in [H–R;II], pp. 263–
290, [D–W], pp. 93–106 and 248–251, and [Pal], pp. 534–538.
The purpose of the present paper is to stress that all these proofs, and also some others, essentially rely upon a lemma which is implicitly con- tained in Cohen’s paper [C]. This lemma appears below as Lemma 1 of Section 2.
Our discussion of various proofs of the Cohen–Hewitt Factorization The- orem is contained in Sections 2 and 3.2–3.4. Section 3.1 concerns the earlier factorization results of R. Salem, A. Zygmund and W. Rudin. In Sections 3.5–3.7 some papers are mentioned containing either more elaborate fac- torization theorems of the Cohen–Hewitt type or factorization results of another nature.
Acknowledgements. The author is greatly indebted to Wojciech Choj- nacki for stimulating discussions and bibliographical hints.
1. The Cohen–Hewitt Factorization Theorem. The left approxi- mate identity in a Banach algebra A is, by definition, a net (e ι ) ι∈I ⊂ A such that
(1) lim
ι ke ι a − ak A = 0 for every a ∈ A.
2000 Mathematics Subject Classification: Primary 46H25; Secondary 43A20.
Key words and phrases : factorization for a representation of a Banach algebra with a bounded approximate identity.
[177]
A left approximate identity (e ι ) ι∈I ⊂ A is said to be bounded if sup
ι∈I
ke ι k A < ∞.
If A is a Banach algebra with bounded left approximate identity (e ι ) ι∈I and T is a continuous representation of A on a Banach space X, then
(2) lim
ι kT (e ι )y − yk X = 0 for every y ∈ span T (A)X.
This follows from the fact that (T (e ι )) ι∈I is a bounded net in L(X) such that lim ι kT (e ι )y − yk = 0 for every y ∈ T (A)X ( 1 ).
The Factorization Theorem (P. Cohen [C], 1959; E. Hewitt [H], 1964). If A is a Banach algebra with bounded left approximate identity (e ι ) ι∈I and T is a continuous representation of A on a Banach space X, then T (A)X is a closed subspace of X. Furthermore, for every y ∈ T (A)X and every ε > 0 there are a ∈ A and x ∈ T (A)y such that T (a)x = y, kx − yk ≤ ε, and
(3) a = X ∞ n=1
p n e ι n where ι n ∈ I, p n > 0 for n = 1, 2, . . . and X ∞ n=1
p n = 1.
The assertions of the Factorization Theorem may be expressed by the single condition
(4) for every y ∈ span T (A)X and every ε > 0 there are a ∈ A and x ∈ T (A)y such that T (a)x = y, kx − yk ≤ ε and (3) is satisfied.
2. Cohen’s proof of the Factorization Theorem. Let A be a Banach algebra, and T a continuous representation of A on a Banach space X. Sup- pose that there is a bounded left approximate identity in A. Then Cohen’s proof of the existence of a factorization
y = T (a)x, a ∈ A, x ∈ X,
of an element y of span T (A)X consists in approximation by a sequence of factorizations
y = e T (a n )x n , a n ∈ G, x n = e T (a −1 n )y, n = 1, 2, . . . ,
where G is the set of invertible elements of the unitization A u of A, and e T is a representation of A u on X extending the representation T .
( 1 ) From (2) it follows at once that span T (A)X = T (A)X, i.e. T (A)X is a closed
linear subspace of X. The factorization theorem says more: T (A)X itself is a closed linear
subspace. This last is almost trivial when A contains a unit e. Indeed, since T (e) =
T (e · e) = T (e)T (e), it follows that T (e) is a projector and hence T (e)X is a closed
subspace. Furthermore, T (A)X = T (eA)X = T (e)T (A)X ⊂ T (e)X ⊂ T (A)X, so that
T (A)X = T (e)X.
The unitization A u of A is the Banach algebra with unit such that 1 o as a linear space, A u is equal to the direct sum K + A, where K is the field of scalars of A,
2 o A u = K + A is equipped with the norm k k A u such that kλ + ak A u =
|λ| + kak A for every λ ∈ K and a ∈ A, 3 o the multiplication in A u is defined by
(λ + a)(µ + b) = λµ + (µa + λb + ab), where λ, µ ∈ K and a, b ∈ A.
The unit in A u is 1 = 1 + 0 ∈ K + A ( 2 ). Let ϕ : A u → K, π : A u → A
be the projectors corresponding to the splitting A u = K + A. From 3 o it fol- lows that ϕ is a multiplicative functional on A u . Let e T be the representation of A u on X such that
T (a) = ϕ(a) + T (πa) e for every a ∈ A u .
The following lemma is implicitly contained in Cohen’s paper [C].
Lemma 1. Under the assumptions of the Factorization Theorem, let E = {e ι : ι ∈ I}, M = sup
e∈E
kek A .
Take any y ∈ span T (A)X, a ∈ G, γ ∈ (0, 1/(M + 1)), δ > 0. Then there exists an ea ∈ G satisfying the three conditions:
ϕ(ea) = ϕ(a) − γϕ(a), (i)
πea = πa + γϕ(a)e for some e ∈ E, (ii)
k e T (ea −1 )y − e T (a −1 )yk X ≤ δ.
(iii)
Note that (i)&(ii) may be written as one condition
(iv) ea = a + γϕ(a)(e − 1),
and that (i)&(ii) implies
(v) kπea − πak A ≤ γM |ϕ(a)| = M |ϕ(ea) − ϕ(a)|.
Lemma 1 implies the Factorization Theorem. Let p n = λ n−1 − λ n for n = 1, 2, . . . ,
where λ 0 , λ 1 , . . . is a positive strictly decreasing sequence such that
(5)
λ 0 = 1, lim
n→∞ λ n = 0, 1 − λ n
λ n−1
= γ n ∈
0, 1
M + 1
for n = 1, 2, . . .
( 2 ) If A has a unit e, then A u makes sense, but e is no longer a unit in A u .
Then (6)
λ n = (1 − γ n )λ n−1 ,
p n = γ n λ n−1 > 0 for n = 1, 2, . . . , and X ∞ n=1
p n = 1.
The concrete form of the sequence λ 0 , λ 1 , . . . is inessential. One can follow [C] and take λ n = (1 − γ) n , γ = const ∈ (0, 1/(M + 1)), so that γ n = γ, or λ n = (1 + M )/(1 + M + n), so that γ n = 1/(1 + M + n). From (5), (6) and (i)&(v)&(iii) it follows that for every y ∈ span T (A)X and every ε > 0 there exists a sequence a 0 , a 1 , . . . of elements of G such that a 0 = 1 and
ϕ(a n ) = λ n , (i) 0
kπa n − πa n−1 k A ≤ M p n , (v) 0
k e T (a −1 n )y − e T (a −1 n−1 )yk X ≤ εp n
(iii) 0
for n = 1, 2, . . . From (i) 0 &(v) 0 &(iii) 0 it follows that lim n→∞ ka n −ak A u = 0 for some a ∈ A such that
kak A = π
X ∞ n=1
(a n − a n−1 ) A ≤ M
X ∞ n=1
p n = M,
and, if x n = e T (a −1 n )y, then lim n→∞ kx n − xk X = 0 for some x ∈ X such that
kx − yk X = kx − x 0 k X ≤ X ∞ n=1
kx n − x n−1 k X ≤ ε X ∞ n=1
p n = ε.
Since y = e T (a n )x n for every n = 0, 1, . . . , a passage to the limit implies that y = e T (a)x = T (a)x. By (2), y = lim ι T (e ι )y ∈ T (A)y, whence
x n = e T (a −1 n )y = ϕ(a −1 n )y + T (πa −1 n )y ∈ T (A)y
for n = 0, 1, . . . , and so x = lim n→∞ x n ∈ T (A)y. Thus (i)&(v)&(iii) implies (4) M for every y ∈ span T (A)X and every ε > 0 there are a ∈ A and
x ∈ T (A)y such that kak A ≤ M , kx − yk X ≤ ε and T (a)x = y.
The statement (4) M is a weakened version of (4): the condition (3) is replaced by a ∈ A, kak A ≤ M . In order to prove (4), instead of (i)&(v)&(iii) one has to use (iv)&(iii). Indeed, (iv)&(iii) implies that for every y ∈ span T (A)X and every ε > 0 there exists a sequence a 0 , a 1 , . . . of elements of G such that a 0 = 1, the conditions (i) 0 and (iii) 0 are satisfied, and
(iv) 0 for every n = 1, 2, . . . there is e n ∈ E such that
a n = a n−1 + γ n ϕ(a n−1 )(e n − 1) = a n−1 + p n (e n − 1).
Since (iv) 0 implies (v) 0 , the preceding argument leading to (4) M remains in
force. Furthermore, from (iv) 0 it follows that a n = λ n + p 1 e 1 + . . . + p n e n for every n = 1, 2, . . . and hence, by (5) and (6), lim n→∞ ka n − ak A u = 0, where a = P ∞
n=1 p n e n .
Proof of Lemma 1 ( 3 ). For every ι ∈ I define ea ι = a + b ι , b ι = γϕ(a)(e ι − 1).
Lemma 1 will follow once it is shown that
(I) there is ι 0 ∈ I such that ea ι ∈ G whenever ι 0 ≺ ι, and
(II) lim
ι k e T (ea −1 ι )y − e T (a −1 )yk X = 0.
In the proof of (I) we will use the equality ea ι = (1 + b ι a −1 )a.
Notice that
b ι a −1 = γϕ(a)(e ι − 1)[ϕ(a −1 ) + πa −1 ]
= γ(e ι − 1) + γϕ(a)(e ι − 1)πa −1 , whence
(7) kb ι a −1 k A u ≤ γ(M + 1) + γ|ϕ(a)| · k(e ι − 1)πa −1 k A . Since γ ∈ (0, 1/(M + 1)), one can choose θ such that
γ(M + 1) < θ < 1.
By (1) and (7), there is ι 0 ∈ I such that
kb ι a −1 k A u ≤ θ whenever ι 0 ≺ ι.
An application of the C. Neumann series shows that then 1 + b ι a −1 ∈ G and k(1 + b ι a −1 ) −1 k A u ≤ (1 − θ) −1 . In consequence, whenever ι 0 ≺ ι, then ea ι = (1 + b ι a −1 )a ∈ G ( 4 ) and
(8) kea −1 ι k A u ≤ ka −1 k A u (1 − θ) −1 . For the proof of (II) observe that if ι 0 ≺ ι, then
ea −1 ι − a −1 = ea −1 ι (a − ea ι )a −1 = −ea −1 ι b ι a −1 = −γϕ(a)ea −1 ι (e ι − 1)a −1 . As a consequence, by (8),
k e T (ea −1 ι )y − e T (a −1 )yk X
≤ γ|ϕ(a)| · k e T k L(A u ;L(X)) ka −1 k A u (1 − θ) −1 k[T (e ι ) − 1] e T (a −1 )yk X .
( 3 ) This proof differs from the argument in [C]; see Section 3.2.
( 4 ) This proof of the invertibility of e a ι follows [P;3], pp. 283–284, and [P–P], p. 136.
The elements b, b ′ and c of A u appearing in [P;3] correspond to our a −1 , e a −1 ι and (1 +
b ι a −1 ) −1 . In [P–P] the elements of A u are written in the form a + r, a ∈ A, r ∈ K, so
that our πa −1 corresponds to (a + r) −1 − r −1 in [P–P].
From this estimate and from (2) the equality (II) follows, because e T (a −1 )y = ϕ(a −1 )y + T (πa −1 )y ∈ span T (A)X.
3. Forerunners of the Cohen–Hewitt Factorization Theorem, variants of the proof, and references to other factorization results 3.1. Forerunners of the Cohen–Hewitt theorem. Let G be a locally compact group with a fixed left-invariant Haar measure µ, and let L 1 (G) be the space of the equivalence classes of functions on G integrable with respect to µ. The convolution a ∗ b of two elements, a and b, of L 1 (G) is defined by
(a ∗ b)(g) =
\
G
a(h)b(h −1 g) µ(dh) for a.e. g ∈ G.
Then L 1 (G) is a convolution Banach algebra. Cohen’s paper [C] contains the following results which follow from the Factorization Theorem applied to A = L 1 (G), X = L 1 (G), or X = C(G) if G is compact, and to the representation of A on X defined by T (a)x = a ∗ x:
(a) L 1 (G) = L 1 (G) ∗ L 1 (G) for every locally compact group G. This was proved earlier in particular cases: for G = T, the circle group, by R. Salem and A. Zygmund, and for G = R and G a locally euclidian abelian group by W. Rudin [R;1], [R;2].
(b) C(G) = L 1 (G) ∗ C(G) for every compact group G. In the particular case of G = T this was proved earlier by R. Salem and A. Zygmund.
(c) Strictly positive factorization: if G is a compact group, y ∈ C(G) and inf G y > 0, then there are a ∈ L 1 (G) and x ∈ C(G) such that ess inf G a > 0, inf G x > 0 and a ∗ x = y. In the deduction of this result from the Fac- torization Theorem the conditions (3) and kx − yk ≤ ε are essential. Note that Cohen [C] proved that an analogous non-negative factorization does not hold.
In their proofs of the particular cases of (a) and (b), R. Salem, A. Zyg- mund and W. Rudin used the Fourier methods which are not applicable in the general situation. Nevertheless it seems interesting to compare the condition (3) from the factorization theorem with the formulas labelled be- low by (∗) and (∗ ∗∗), appearing in the proof of the factorization L 1 (R) = L 1 (R) ∗ L 1 (R) presented in [R;2]. The approximate identity in L 1 (R) ap- pears there in the from e t = K t for t ∈ (0, ∞), where
K t (x) = t 2π
sin(tx/2) tx/2
2
is the Fej´er kernel. The Fourier transform of K t is
K b t (ξ) = (1 − |ξ|/t) + .
For given f ∈ L 1 (R) Rudin proves that f = a ∗ g where a ∈ L 1 (R) and g ∈ L 1 (R) are defined as the integrals of continuous L 1 (R)-valued functions
a =
∞
\
0
tλ ′′ (t)K t dt, (∗)
g = f +
∞
\
0
α(t)(f − K t ∗ f ) dt.
(∗ ∗)
Here α ∈ C[0, ∞) and λ ∈ C 2 [0, ∞) are positive functions such that
∞
\
0
α(t) dt = ∞,
∞
\
0
α(t)
kK t ∗ f − f k + 1 t
dt < ∞, λ(t) = 1 Φ(t) , where
Φ(t) = 1 +
t
\
0
α(s) ds + t
∞
\
t
α(s) s ds.
These definitions imply that g ∈ L 1 (R), λ ′′ is positive and (∗ ∗∗)
∞
\
0
tλ ′′ (t) dt = 1,
so that a ∈ L 1 (R), and furthermore, ba(ξ) = λ(|ξ|) and bg(ξ) = Φ(|ξ|) b f (ξ) ( 5 ), whence babg = b f , and so a ∗ g = f .
Discrete forerunners of (∗) and (∗ ∗∗) appear in the book of A. Zygmund [Z;I] as (1.7) on p. 183, (4.2) on p. 93, and Theorem 1.5 on p. 183 (attributed on p. 378 to H. W. Young (1913) and A. N. Kolmogorov (1923)). A discrete forerunner of (∗ ∗) is the second formula in line 111 3 of R. Salem’s paper [S], related to a limit passage in (3) on p. 110 of [S]. A. Zygmund [Z;I], p. 378 11-6 , points out that the theorems from [S] together with Theorem 1.5 of Chap. V of [Z;I], p. 183, lead to the factorization results for G = T mentioned in (a) and (b). A discrete analogue of the whole elegant Rudin construction is used in Sec. 7.5 of the book of R. E. Edwards [E;I] for reproving the Salem–
Zygmund results. Remarks on the Fourier factorization may be found in Sec. 3.1.1(c) and 7.5.2–4 of [E;I] and Sec. 11.4.18(6) of [E;II].
( 5 ) Indeed, Φ ′ (t) =
T
∞
t (α(s)/s) ds, Φ ′′ (t) = −α(t)/t ≤ 0, λ ′ (t) = −Φ ′ (t)/[Φ(t)] 2 ≤ 0, λ ′′ (t) = (2[Φ ′ (t)] 2 − Φ ′′ (t)Φ(t))[Φ(t)] −3 ≥ 0. Hence 0 ≤ −tλ ′ (t) = tΦ ′ (t)/[Φ(t)] 2 ≤ [Φ(t)] −1 , so lim t→∞ tλ ′ (t) = lim t→∞ λ(t) = 0 and
T
∞
0 tλ ′′ (t) dt = tλ ′ (t)| t=∞ t=0 −
T
∞ 0 λ ′ (t) dt
= λ(0) = 1. Moreover, (∗) implies b a(ξ) =
T
∞
0 tλ ′′ (t)(1 − |ξ|/t) + dt =
T
∞
|ξ| (t − |ξ|)λ ′′ (t) dt = (t − |ξ|)λ ′ (t)| t=∞ t=|ξ| −
T
∞
|ξ| λ ′ (t) dt = λ(|ξ|), and bg(ξ) = [1 +
T